FMA ‐ DETR : A Frequency‐Enhanced Multi‐Scale Attention Transformer for Small Object Detection in UAV Aerial Imagery

ABSTRACT Object detection in drone aerial imagery finds extensive applications in emergency response and autonomous surveillance. However, constrained by high‐altitude perspectives, small targets exhibit blurred features, drastic scale variations, and susceptibility to background interference. To address these challenges, this paper proposes the Frequency‐domain Enhanced Multi‐scale Small Object Detection for Drone Aerial Imagery (FMA‐DETR). Specifically, first, a multi‐branch frequency‐domain parallel attention C2f module (C2f‐MFPA) is designed within the backbone network. This module integrates local–global attention and frequency‐domain gating paths through a four‐branch parallel architecture, enhancing fine‐grained feature extraction while effectively reducing channel redundancy. Second, a frequency‐domain enhanced feedforward network module (EDFFN) is introduced into the encoder, employing a patch‐wise FFT strategy to accentuate edge texture features of targets. Finally, a multi‐scale convolutional attention feature pyramid module (MSCA‐FPN) is proposed, achieving adaptive integration of cross‐level features through a multi‐scale feature alignment module (Zoom_cat) and a dual‐path fusion mechanism (CAFMFusion). Experimental results on the VisDrone2019 and HIT‐UAV datasets demonstrate that compared to baseline models, our approach achieves a 3.1% improvement on the VisDrone2019 dataset while maintaining the same parameter count. On the HIT‐UAV dataset, the metric improved by 1.4%. These results highlight an excellent balance between model capacity and detection performance, providing an efficient solution for edge deployment on UAVs.